MFM Practitioner Module: Risk & Asset Allocation. John Dodson. January 25, 2012

Transcription

1 MFM Practitioner Module: Risk & Asset Allocation January 25, 2012

2

3 Optimizing Allocations Once we have 1. chosen the markets and an investment horizon 2. modeled the markets 3. agreed on an objective with the client 4. designed an index of satisfaction We are almost ready to determine the optimal portfolio The last item we need to consider before proceeding with the mechanical exercise of optimization is the domain of possible allocations, C Wealth Constraint One constraint that almost always applies is the wealth constraint C = {α p T α = w T } for given initial wealth w T.

4 Secondary In addition to or instead of the wealth constraint, other constraints may be relevant transactions costs exogenous risk measures and limits prohibition against short positions Each binding constraint has a corresponding shadow cost reflecting the fact that the otherwise unconstrained optimal portfolio is not feasible. Professional investment management is largely about managing constraints. N.B.: The form and complexity of the constraints determines the form and complexity of the optimization techniques required.

5 Problem The optimal portfolio, if it exists, is defined by α = arg max α C S(α) Constrained optimization in general is a hard problem. There are several classes that are more tractable. Convex Programming If the objective Q is convex, and the constraints can be represented by the intersection of sets L = {z Az = a} and V = {z F (z) 0} for convex F ( ), the problem arg min z L V Q(z) has a solution, although it may be difficult or expensive to determine. This is another reason we should choose an index of satisfaction that is concave.

6 If the index of satisfaction is consistent with stochastic dominance, we may be able to replace the original hard problem with two simpler problems. Stochastic Dominance An allocation α is said to weakly dominate an allocation β if F Ψα (ψ) F Ψβ (ψ) ψ in support and an index of satisfaction is consistent with weak stochastic dominance if S(α) S(β). value-at-risk and expected shortfall are both consistent with weak stochastic dominance In particular, if Sd Ψ α = Sd Ψ β but E Ψ α > E Ψ β then S(α) > S(β)

7 Two-Step Procedure As long as the index of satisfaction is consistent with weak stochastic dominance, we can exclude from consideration in the search for the optimal portfolio any portfolio that is not on the efficient frontier, defined by 1. α (v) = arg max α C E Ψ α var Ψ α v The optimal portfolio is therefore the solution to a one-dimensional unconstrained problem, 2. v = arg max v S (α (v)) which can be found numerically, viz. α = α (v ). This framework presents an enormous computational improvement, and justifies choosing an index of satisfaction that is consistent with weak stochastic dominance.

8 For a single, affine constraint, such as the wealth constraint, the efficient frontier can be derived analytically. First we introduce the dual problem αdual (e) = arg min d α=c α E M e α (cov M) α The constraints can be handled by introducing two (scalar) Lagrange multipliers, λ and µ. min α,λ,µ α (cov M) α λ ( α d c ) µ ( α E M e ) Any solution to this for c > 0 (see Appendix 6.3) can be expressed as a linear combination of two portfolios, α SR = c (cov M) 1 E M d (cov M) 1 E M α MV = c (cov M) 1 d d (cov M) 1 d

9 The efficient frontier consists of all portfolios α MV + β (α SR α MV ) β 0 N.B.: These portfolios are orthogonal, in the sense that cov (Ψ αsr Ψ αmv, Ψ αmv ) = 0 Sharpe Ratio Portfolio The portfolio α SR is the unique portfolio that maximizes the Sharpe Ratio. α SR = arg max d α=c E Ψ α Sd Ψ α Minimum Variance Portfolio The portfolio α MV is the globally minimum variance portfolio within the feasible set. If there is an investment that is risk-free relative to the objective, e.g. cash or a benchmark-neutral allocation, then it will consist of this.

10 The first step of the two-step procedure requires the mean and covariance of the market vector, regardless of the form of the index of satisfaction. Therefore, let us review these quantities for the standard log-normal model P T +τ = p T e X with X N (µ, Σ) where the exponentiation and is component-wise. Recalling that φ X (t) = e iµ t 1 2 t Σt, we can get Mean E P T +τ = p T e µ+ 1 2 diag Σ Covariance cov P T +τ = (E P T +τ ) (E P T +τ ) (e Σ 11 )

11 The efficient frontier depends in a non-trivial way upon the investment horizon, τ. Say you had a model for the invariants X N (µ, Σ) based on a sampling period of τ, but you were interested in a different investment horizon τ. We know that X is additive, while e X is not. In particular, µ+ 1 2 diag Σ E P T + τ = p T e τ τ cov P T + τ = (E P T + τ ) (E P T + τ ) (e τ τ Σ 11 ) which leads to a different solutions for α SR and maybe α MV. This is not apparent in the traditional analysis in terms of continuous returns and underscores the importance of choosing an appropriate investment horizon.

12 When the investor s objective is relative, the wealth constraint is of the form d α = 0. In this case, the minimum variance portfolio is trivial, α MV = 0; but the solution for the maximum Sharpe Ratio portfolio fails (since c = 0). In place, we can use α SR = (cov M) 1 ( ) d (cov M) 1 d E M d (cov M) 1 (E M) d and all efficient portfolios have the same Sharpe Ratio (termed the Information Ratio in the relative setting). Implied Benchmark The efficient frontier for the absolute objective is the same as the efficient frontier for the objective relative to the global minimum-variance portfolio α MV.